SWE-AF:基于 AgentField 的开源自治工程团队运行时
SWE-AF 在 AgentField 上把一个工程目标拆成计划、编码、评审、测试与校验的多智能体流水线,支持单仓/多仓、按角色选模型、并行 worktree 与断点续跑,Apache 2.0。
单个编程智能体可以改文件;要把“自然语言目标”变成可审查、可合并的变更,还需要计划、分工、失败回退和仓库治理。SWE-AF(读作 swee-AF)把自己定位为跑在 AgentField 上的自治工程团队运行时:一次调用拉起产品、架构、编码、评审、测试等角色,围绕共享构建流协作。
帖文强调的能力包括:单仓或多仓模式(可区分主仓与依赖仓);按 coder、QA、architect 等角色配置运行时与模型;QA 或评审失败时进入重试、议题级调整或对其余工作重规划;依赖级调度配合隔离 Git worktree,降低并行改分支冲突;以及检查点续跑,崩溃后可恢复。许可证 Apache 2.0,仓库为 github.com/Agent-Field/SWE-AF。
公开 README 把入口写成 af install 后调用 swe-planner.build,并支持 Claude Code、OpenCode、Codex 等运行时。对已在 AgentField 控制面运行节点的团队,它提供的是可安装的“工程工厂”节点,而不是再手写一套编排。公开材料中的工期对比属于项目方示意,落地效果仍取决于目标难度、模型与仓库约束,需自行验证。
flowchart TD
goal[工程目标] --> plan[规划与议题 DAG]
plan --> exec[并行 worktree 执行]
exec --> qa{QA或评审}
qa -->|失败| adapt[重试或重规划]
adapt --> exec
qa -->|通过| verify[验收与修复环]
verify --> pr[推送草稿 PR]
整体上,它把多智能体工程协作从演示脚本推进到可安装运行时;是否适合你的仓库规模与合规要求,仍要以小范围试点结果为准。
SWE-AF autonomous engineering-team runtime
A coding agent can write a file. This repo coordinates the engineering work around it.
SWE-AF is an autonomous engineering-team runtime built on AgentField for builders running software work against one or multiple repositories.
It helps you turn a defined engineering goal into planned, coded, reviewed, tested, and verified work by coordinating specialized agents through a shared build flow.
Key features:
• Single- and multi-repo modes – run a build against one local or remote repo, or assign primary and dependency roles across repos • Role-based model configuration – choose a runtime and set models for roles such as coder, QA, and architect • Adaptive control loops – QA or review failures can trigger retries, issue-level adaptation, or replanning of the remaining work • Parallel worktrees – dependency-level scheduling and isolated Git worktrees support concurrent issue work without branch collisions • Resumable execution – checkpointed builds can resume after crashes or interruptions
It’s open-source (Apache License 2.0 license).
Link in the reply 👇